Multi-object Tracking In Video Based On Deep Learning

Multi-object tracking in video is an important fundamental problem in computer vision.It has various applications such as intelligent video surveillance,autonomous driving,intelligent robot,intelligent human-computer interaction,sports video analysis and so on.Frequent occlusion between multiple targets makes it difficult for extending single object tracking algorithm to multi-object tracking task.The tracking efficiency of single object tracking algorithm decreases dramatically with the number of tracked targets increasing,which further limits its application in multi-object tracking scenar-ios.In addition,the existing multi-object tracking algorithms based on data association heavily depend on the performance of object detector and are unable to utilize the re-lationship between object detection and tracking in video.This dissertation focuses on developing effective algorithms to handle these issues.The main contributions and innovations are as follows:1.Proposed an online multi-object tracking algorithm based on anti-occlusion single object tracking model.The algorithm uses spatio-temporal attention mech-anism and adversarial training to improve the robustness of tracking model to occlusion,which solves the tracking drift problem caused by occlusion and ef-fectively extends the single object tracking algorithm based on deep learning to multi-object tracking task.The spatial attention mechanism weights the feature map of the target spatially so that the features of the un-occluded area get more attention.Thus,the tracking result is more accurate when the target is occluded.The temporal attention mechanism controls the importance of training samples with different occlusion statuses in different frames.Thus,the more severely a sample is occluded,the smaller the impact on model updating is.In order to make up the lack of occluded samples collected online,adversarial training is used to generate occluded samples for updating tracking model.The experimental re-sults demonstrate that the proposed online multi-object tracking algorithm can effectively alleviate the tracking drift problem caused by occlusion.2.Proposed an online multi-object tracking algorithm combining single object tracking and data association.The algorithm combines the advantages of data association based tracking and single object tracking.It adopts the Siamese con-volutional neural network to fuse them into a unified network.Through feature sharing and interaction during training,data association and single object tracking are learned to complement each other.In addition,in order to handle the problem that the tracking efficiency of single object tracking algorithm extended to multi-object tracking scenarios decreases greatly with the number of tracked targets increasing,an efficient two-stage single object tracking algorithm is proposed,which utilizes the merits of correlation features and can simultaneously track all the existing targets within one forward propagation.The experimental results show that the fusion of data association and single object tracking can effectively improve the tracking performance.The efficient two-stage single object tracking algorithm makes the tracking efficiency independent of the number of tracked targets.Compared with the existing online multi-object tracking methods,the proposed algorithm strikes a better balance between tracking performance and tracking speed.3.Proposed a joint online multi-object detection and tracking algorithm.The algorithm combines video object detection and tracking into a unified network,which solves the problem that the existing multi-object tracking algorithm regards object detection and tracking as two independent tasks and cannot utilize the rela-tion between them.During offline training,in order to ensure the stability of the object detector,a novel loss function is designed to penalize the difference in the detection scores of the same target in different frames.During the online track-ing process,the detection result is refined according to the tracking result,and the tracking result is fed back to the object detector.The object detector is fine-tuned online,so that it can adapt to the current video scenario to obtain higher detection performance.The tracking performance is further improved in turn.The exper-imental results show that the proposed algorithm can improve the performance of video object detection and tracking at the same time,which demonstrates the effectiveness of joint online multi-object detection and tracking.